Guide

Meta Muse, Wonder and Spark Open: Indonesia Access and Costs

Updated 2026-10-16 · by Fahmi Fahreza

Muse Spark Open adds downloadable weights and hosted-provider access, with limited commercial license terms and substantial hardware needs. Keep it separate from Wonder, Glimmer, and personal Muse agent eligibility when planning an Indonesian pilot.

What changed with Muse Spark Open?

Meta released Muse Spark Open on October 15, 2026, with downloadable weights and hosted-provider options. Its launch announcement specifies the Muse Model License with limited commercial terms. Open-weight does not mean unrestricted open-source or free infrastructure. For an Indonesian team, first choose between a bounded hosted pilot and operating the model yourself; neither establishes access to the consumer Muse agent.

Where does Muse Wonder fit?

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026. Its official announcement says Wonder is available through Meta Model API and Muse Code. Its rollout in the personal agent Muse is still upcoming, starting with subscribers. The model release does not confirm Muse agent access for Indonesian accounts.

The Spark Open update was checked on October 16, 2026; individual source-check dates are recorded separately. The examples below are proposed exercises and hypothetical calculations, not hands-on test results, evidence of Indonesian-language accuracy, or proven sales improvements.

Distinguish the Muse agent, Wonder model, and Muse Code

Muse is a personal AI agent with a cloud computer and browser, as described in its product design explanation. Model API supplies models for applications you build, while Muse Code is a terminal coding agent. Meta's developer overview labels Wonder a preview, with model ID muse-wonder-1.0 and a 1,048,576-token context window.

This access map summarizes Meta's documentation, rather than verification of an Indonesian account:

Access routeModel and statusWhat to check
Meta Model APIWonder; pay-as-you-go, Standard modeAccount eligibility, usage limits, budget
Muse CodeWonder; pay-as-you-go or every subscription planCLI access and account plan
Muse Code for FullFull; Power Usage or Maximum Usage onlyCorrect plan and data settings
Personal agent MuseWonder coming soon, subscribers firstFeature rollout in participants' accounts

The first three rows follow Meta's model catalog; the last follows the Wonder announcement. Full is unavailable through the pay-as-you-go API, and max reasoning on Wonder does not select Full. Wonder is a hosted model; do not treat it as an open-weight release. The developer overview now distinguishes Spark Open, available through self-hosting or inference providers, from Muse Glimmer, designed for local and edge use under Apache 2.0. Spark Open is not served through Meta Model API; the hosted Muse Spark API model is a separate access route.

A Muse Code subscription covers its dedicated CLI credential; additional API keys are billed pay-as-you-go. Subscription benefits and availability can vary by region. Check onboarding details before paying. Source: Muse Code subscriptions.

Spark Open: hosted pilot or self-hosting?

A hosted provider can avoid buying hardware for a first evaluation. Check its model identifier, Indonesian account eligibility, supported inputs and tool calls, usage limits, data retention, and current charges. Do not apply the Wonder API prices below to Spark Open. Meta's release announcement lists provider options, but this guide does not verify any provider account or recommend one on measured performance.

Self-hosting needs a concrete capacity plan. Meta's hardware guide documents a four-GPU NVIDIA Blackwell configuration for the NVFP4 checkpoint. Its Apple route uses MLX across two M5 Ultra Mac Studios with 512 GB unified memory each, or four with 256 GB each, with RDMA over Thunderbolt 5, macOS 26.2 or later, and direct links between every pair. These are documented configurations, not a claim that an ordinary laptop can run the full model.

Before buying or renting hardware, validate the exact checkpoint, runtime, memory, context length, concurrency, and networking requirements. Budget for machine rental or purchase, power, storage, monitoring, updates, security, and staff time. Review the license linked from the official model release for your intended use; downloadable weights alone do not establish commercial permission.

For a small team, start with fictional data and a fixed task budget on an approved hosted service if its terms fit. Consider self-hosting only when control needs, sustained workload, and operational capacity justify it. This is a planning recommendation, not a measured cost comparison. Self-hosting also does not prevent data leaving through connected tools, logs, or external services.

Indonesia access and business needs

The reviewed sources do not confirm personal agent Muse access in Indonesia. The Muse for Small Business announcement, updated October 8, still names the US and Canada. Assess that coverage separately from API and Muse Code eligibility.

Meta lists professional Instagram accounts, Facebook Pages, Canva, and Shopify connections in that business announcement. Inspect participants' connectors before scheduling training. An integration name does not establish that the entire business process will work.

If access is unavailable, practice briefing, permission boundaries, and evaluation with sample documents. Do not promise live exercises using features that have not appeared in participants' accounts.

What does the Muse Wonder API cost?

Meta's listed prices for muse-wonder-1.0 in Standard mode, in USD per million tokens:

  • Input: $4.00
  • Cached input: $0.20
  • Output: $20.00

There is no long-context premium. Web search adds $2.50 per 1,000 queries beyond token charges. These are API usage rates, not subscription prices or guaranteed task costs.

Hypothetical example: a total of 100,000 uncached input tokens and 10,000 output tokens costs $0.60 in tokens: $0.40 + $0.20. Twenty web searches add $0.05, giving a $0.65 subtotal. This excludes taxes, currency conversion, other services, and infrastructure. It is not a bill from an actual test.

For a pilot, count every call, including sub-agents, retries, and revisions. Set a spending cap and stopping conditions before a run; a large context window is not a reason to fill it.

Check privacy before using client code or data

Wonder on the API supports only Standard mode, whose content is not used to train Meta models. Source: model modes and availability.

Muse Code subscriptions enable Contributor mode by default. It grants more usage in exchange for allowing content to train Meta models. Turning it off excludes future content from training while it remains off. Content includes prompts, code, files, context, and responses. Source: Contributor settings.

Do not extend those settings to the personal Muse agent. Its security explanation documents an available training opt-out in Muse settings. That explanation still describes Confidential VM as planned. Do not count promised protection as active or interpret “not used for training” as a guarantee of no storage.

Before opening a client repository or document, check the data owner's permission, organizational policy, the correct product's settings, and third-party services receiving data. Start with a sample repository, public catalog, or fictional data; remove API keys, passwords, and customer information.

Proposed pilot: prepare next week's promotion

Choose a sample business such as a packaged-food shop. The target is narrow: three content drafts for the owner to review, without publication. An education provider could substitute public course information for the catalog. Use an available, organization-approved tool; this exercise does not require Wonder Full.

Prepare a small source packet:

  • A public catalog showing prices, units, validity dates, and source links.
  • Five fictional customer questions, including one with no available answer.
  • Approved writing examples without customer names or phone numbers.
  • Prohibited claims, such as health benefits or unapproved discounts.
  • A reviewer responsible for checking the output before use.

Try this exercise brief:

Help prepare next week's promotion from the attached catalog. Produce three Indonesian-language caption drafts and a list of missing information. Cite sources for prices and availability. Mark missing information as unknown. Do not invent discounts, publish content, contact anyone, or change accounts. Return the drafts only in this conversation for my review.

Separate factual checks from writing feedback. An attractive caption still fails if it confuses a pack price with a unit price. Ask the agent to flag questions requiring an owner's decision, then improve the source packet before trying again.

If using the API in a workflow, implement permissions and review in your own integration. Changing the model does not automatically supply every control in the personal Muse agent product.

Action controls and success criteria

For the personal Muse agent, its product design describes activity logs and approval cards. The connector guidelines require approval for every sensitive change and prohibit setting those actions to Always allow. Choose read-only access where available for an initial exercise.

Meta acknowledges that agents can still make mistakes. Set the evaluation rubric before seeing the output:

  • Record preparation and review time; compare with manual work using the same inputs.
  • Count factual corrections, citations that do not support claims, and instruction violations.
  • Test conflicting prices, unknown stock, and a discount request embedded in a sample question.
  • Record all costs and cases that need human handling.

Acceptable output should expose uncertainty and request a decision. If the agent invents promises, stop expanding access and repeat with fictional data. Add one source or permission at a time so failures remain traceable.

Distinguish it from the former video service

The video platform previously called muse.ai announced its Skiv name on March 3, 2026. Skiv's announcement covers the name and embed-address changes. Meta's announcement links to muse.ai as the agent's website. Follow Skiv's instructions for legacy video accounts; video-hosting reviews do not establish the Meta agent's capabilities.

Prepare a training exercise with Fahmi

For context on agent choices, read OpenAI Dots vs Meta Muse. Teams looking to practice delegation and output review can explore AI Automation Trainer Indonesia or contact Fahmi. Bring one workflow, safe examples, a spending limit, success criteria, and a hosted-versus-self-hosted decision to discuss; discuss learning needs and participant access before scheduling practice.

Frequently asked questions

Is Muse Spark Open free to use on an ordinary laptop?

Open weights do not mean free infrastructure or unrestricted use. Meta documents substantial multi-device hardware configurations, and the Muse Model License has limited commercial terms. Hosted providers have their own charges and conditions.

Is Meta Muse the same as the muse.ai video platform?

They are different products. The video platform formerly called muse.ai announced its Skiv name on March 3, 2026. Meta uses muse.ai for its personal AI agent Muse.

Do Wonder or Spark Open releases confirm Muse access in Indonesia?

No. Model releases do not establish personal Muse agent eligibility in Indonesia. Check agent, Meta Model API, Muse Code, and third-party provider access separately.

Does max reasoning unlock Muse Wonder Full in the API?

No. Full is available only in Muse Code through Power Usage or Maximum Usage plans; max reasoning on Wonder does not select Full.

What is a sensible first pilot for a small business?

Start with promotion drafts based on a public catalog and fictional questions. Record review time, factual corrections, task-boundary compliance, and costs before expanding access.

Want a session like this for your team or event?

Share your training needs through the form. Fahmi will contact you by email or WhatsApp.